3 ms·

You burn tokens with an smaller LLM you don't care about that exclusively does tool selection (or routing?)